(Yicai) Sept. 24 -- An industry forum has heard how the rapid growth of artificial intelligence has completely changed the field of cybersecurity, with new strategies needed to protect against risks.

The emergence of large language models and AI agent technology is overturning and reshaping the traditional paradigm of cybersecurity defense, speakers said at the forum during the fifth Global Digital Trade Expo, which opened yesterday in Hangzhou.

Security incidents involving AI models have made the headlines in recent months, showing how the risks are rapidly spreading. For example, US firm OpenAI’s agent hacked into Hugging Face, an open-source platform for AI models and datasets. And Claude, a model developed by AI firm Anthropic, gained unauthorized access to three institutional systems during testing.

According to an AI industry report released by an overseas institution this month, 89.5 percent of the surveyed organizations have been hacked by generative AI in the past 12 months, surging upward from 75.1 percent in 2025. Also, 88.4 percent of the surveyed organizations reported at least one hacking incident related to AI agents.

Several guests said at the forum that the current cybersecurity confrontation mode has shifted from human-centered to agent-based, with agents making autonomous decisions. Meanwhile, the intervention window for humans in human-computer interaction is narrowing, and humans may not have time to understand the situation, let alone intervene, which multiplies the risks.

“The endogenous risk of the system has exceeded that of external attack,” said Wu Shizhong, a member of the Chinese Academy of Engineering. “Now the biggest danger is no longer just hacking; it also encompasses the inherent unexplainability, biases, and hallucinations of algorithms, as well as unauthorized actions triggered by manipulation -- malicious behavior, essentially.”

The security game in the era of AI has escalated into a competition between “people who have mastered AI” in the digital space, Wu said, adding that the key factor is how to calibrate and align the system.

“In terms of time, we must shift from post-incident response to prediction within a near-instantaneous timeframe,” Wu stated. Within a millisecond-level window, AI must accurately detect highly covert precursors to attacks amidst massive amounts of network noise.

It’s also necessary to predict and intercept advanced sustainable threats before they form, which will break the cycle of lag inherent in the traditional "patch-after-attack" approach, he noted.

Wang Xin, director of the DAS-Security Research Institute, told Yicai that the development of AI models in the past few years has broken the “balance of attack and defense.” This has prompted enterprises to attach greater importance to AI security issues.

The security business has mainly focused on human-centered detection and defense thus far, but it may shift to being AI-centered in the future, Wang noted, adding that the demand for AI security is growing much faster compared to traditional cybersecurity, and many companies are investing in this field.

According to recent International Data Corporation forecasts, China’s AI security revenue will jump from CNY4.41 billion (USD657.57 million) in 2025 to CNY34.03 billion in 2030, with a five-year compound annual growth rate of about 50.5 percent.

In contrast, the estimated compound annual growth rate of the country’s cybersecurity market from 2024 to 2029 is only about 8.9 percent, with the market expected to grow to about CNY80 billion in 2026.

Editor: Tom Litting